System for marine ecological monitoring information compression and data transmission

Through federated learning and dynamic differential compression coding, the communication resources and bandwidth problems of deep learning model training and information transmission in the marine environment are solved, and efficient monitoring and transmission of marine biological information is realized.

CN119052410APending Publication Date: 2024-11-29NAVAL UNIV OF ENG PLA
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410934814.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

In the marine environment, the deep learning model training and information transmission of drones in the prior art for monitoring marine organisms has problems such as large communication resource occupancy, high bandwidth consumption and long transmission distance, which makes it difficult to effectively monitor and transmit information of marine organisms.

Method used

The distributed training scheme of deep learning model with federated learning is adopted, combining feature dynamic differential compression coding and roaring bitmap compression coding, and the asynchronous update of the drone model and efficient transmission of information through the delay tolerance relay protocol.

Benefits of technology

It reduces the communication bandwidth consumption of drone monitoring information, improves the transmission efficiency of marine biological monitoring data, and ensures effective monitoring and transmission of marine biological information in distant sea environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119052410A_ABST
    Figure CN119052410A_ABST
Patent Text Reader

Abstract

The invention discloses a system for marine ecological monitoring information compression and data transmission. Comprising a plurality of unmanned aerial vehicles, a plurality of buoys deployed in different sea areas, a target marine organism video acquisition module, a deep learning model training module, a marine organism attribute acquisition module, a marine organism attribute coding module, a roaring bitmap compression coding module, a roaring bitmap compression coding sending module and an unmanned aerial vehicle alternating module. According to the system, a deep learning model distributed training scheme based on federal learning is proposed, and deep learning models carried on all unmanned aerial vehicles are asynchronously updated; and characteristic dynamic differential compression coding and roaring map compression coding are designed to compress the monitoring data, so that the coding length and the sending frequency of the monitoring data are reduced, the bandwidth occupancy rate is reduced, and the communication efficiency is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of marine ecological monitoring, and in particular to a system for marine ecological monitoring information compression and data transmission. Background Art

[0002] As people pay more and more attention to the protection of marine ecological environment, marine life is an important part of the marine ecological environment. By monitoring the population, distribution and behavior of offshore organisms, we can understand the health of the ecosystem, thereby providing a scientific basis for the protection and management of marine biological resources. However, due to the complexity of the marine environment, it is difficult to effectively deploy monitoring equipment. Secondly, due to the migration of marine organisms, ordinary static monitoring equipment is difficult to continuously monitor marine organisms. Nowadays, drone technology is widely used. Due to the generally small size of drones, they are easy to be deployed in ocean buoys. In addition, drones have strong maneuverability and flexibility, can quickly adjust the flight trajectory, and can continuously track and monitor marine organisms in motion. In order to transmit the relevant information of the target marine organisms monitored by drones to the monitoring center, it is necessary to use artificial intelligence technology to infer the category, movement status and other information of marine organisms through the trained deep learning model. In addition, in order to ensure that the marine biological information collected by the drone can be successfully received and understood by the monitoring center, it is necessary to establish a communication link between the drone and the monitoring center, and design the corresponding frame format and related communication protocol.

[0003] However, in the marine environment, due to the limitations of communication distance and communication resources, the deep learning model training algorithm and information transmission in the existing technology have the following shortcomings: First, all model training data is collected and stored by drones. If the training data needs to be transmitted to a unified location for centralized training, it will take up a lot of communication resources. Secondly, the direct encoding and transmission of the monitoring results of marine life by drones will occupy a large communication bandwidth and reduce communication efficiency. Finally, drones usually monitor target marine life in the open sea, and the distance between them and the monitoring center is far beyond the coverage of wireless communication. The relevant detection results obtained by drones are difficult to transmit to the monitoring center.

[0004] Hydroacoustic link transmission is a commonly used communication method in the ocean. Compared with wireless link transmission, it has a longer transmission distance and is suitable for long-distance transmission in the ocean. However, the relevant monitoring results collected by drones cannot be directly transmitted through the hydroacoustic link. They must first be transmitted to the ocean buoy through a wireless link, and then forwarded to the monitoring center by the ocean buoy through the hydroacoustic link.

[0005] Therefore, how to design a model training and information transmission method that can save communication resources to overcome the weak connection environment in the ocean is an urgent problem to be solved in the field of marine biological monitoring. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art. The present invention proposes a system for marine ecological monitoring information compression and data transmission, which is used to complete the deep learning model training and transmission of drone monitoring information deployed in all drones with minimal communication consumption.

[0007] The technical idea of ​​the present invention is: first, a distributed training scheme for deep learning models based on federated learning is designed to train different deep learning models carried on drones; secondly, feature dynamic differential compression coding and roaring bitmap compression coding are designed to reduce the coding length and sending frequency of monitoring data.

[0008] Based on the above technical ideas, the present invention provides a system for marine ecological monitoring information compression and data transmission, multiple drones and multiple buoys deployed in different sea areas, and the system also includes:

[0009] A target marine life video acquisition module is used to acquire a video stream of the target marine life through an optical device on a group of drones;

[0010] A deep learning model training module, which is used to update the deep learning models on different drones using a federated asynchronous training method;

[0011] The marine life attribute acquisition module is used to obtain the attribute values ​​of the target marine life through different deep learning models on the drone;

[0012] The marine life attribute coding module is used to compare the current value of each marine life attribute with the attribute value obtained last time. If there is a change, the change value of the attribute is subjected to characteristic dynamic differential compression coding;

[0013] The roaring bitmap compression coding module is used to perform roaring bitmap compression on multiple feature dynamic differential compression codes stored in the same queue to form a message and send it to the roaring bitmap compression coding sending module;

[0014] The roaring bitmap compression encoding sending module is used to send the received message to the buoy to which the UAV is to return;

[0015] The drone alternation module is used to replace the current drone with a new one when the drone currently performing the mission needs to return for charging, and continue to perform the monitoring mission of the target marine life.

[0016] Furthermore, the federated asynchronous training method in the deep learning model training module includes the following steps:

[0017] Step 1: Each drone uses the local dataset to perform mutual distillation training on the deep learning model it carries and the local teacher model corresponding to the deep learning model;

[0018] Step 2: Each UAV obtains the communication status of the current sea area and asynchronously sends the parameters of each model carried by the UAV to the buoy to which it is returning;

[0019] Step 3: The global model generator deployed in the buoy aggregates the received model parameters at fixed time intervals to generate a new global model;

[0020] Step 4: When the UAV returns to the corresponding buoy, the buoy sends the updated global model to the UAV, and the UAV uses the global model to update the local teacher model.

[0021] Furthermore, the attributes of the target marine life include the category of the target marine life, the movement speed of the target marine life, the current position of the target marine life, and the movement direction of the target marine life.

[0022] Further, the marine life attribute coding module includes an attribute change comparison unit and a characteristic dynamic difference compression coding unit;

[0023] The attribute change comparison unit is used to compare the attribute values ​​of the current target marine organisms with the corresponding attribute values ​​last sent to the buoy. If the attribute change value exceeds the preset threshold value of the attribute, the attribute change value is sent to the characteristic dynamic differential compression coding unit.

[0024] The characteristic dynamic differential compression coding unit is used to perform binary coding on the change value of the attribute and add index coding to obtain the characteristic dynamic differential compression coding.

[0025] Furthermore, the frame format of the characteristic dynamic differential compression encoding includes a drone number field, a drone position change field, a marine life number field, a species category change field, a current position change field, a movement speed change field, a movement direction change field and a timestamp field in sequence, and each attribute name represented in the encoding is identified through an index code table.

[0026] Furthermore, the steps of the roaring bitmap compression coding module for roaring bitmap compression coding are: first extract the coding segments with the same high bits in the characteristic dynamic differential compression coding of each attribute, and add the corresponding low bits and the number of low bits and the attribute name coding after the coding segment to form the final roaring bitmap compression coding.

[0027] Furthermore, the specific steps of the drone alternation module using a new drone to replace the current drone to continue the mission are:

[0028] Step 1: Compress all the feature dynamic differential compression codes cached by the returning drone into a roaring bitmap to obtain a roaring bitmap compression code;

[0029] Step 2: The returning drone sends the roar bitmap compression code to buoy A where the returning drone is located through a wireless link;

[0030] Step 3: Buoy A decodes the received roar bitmap code, converts the code into an underwater acoustic link code through a delay-tolerant relay protocol, and then sends it to buoy B where the UAV that is about to take over the mission is located through the underwater acoustic link;

[0031] Step 4: Buoy B decodes the received roar bitmap code, converts the code into a wireless link code through a delay-tolerant relay protocol, and then sends it to the new UAV;

[0032] Step 5: The new drone decodes and restores the data, thereby replacing the returning drone to continue the monitoring mission of the target marine life.

[0033] Furthermore, the delay-tolerant relay protocol includes the following steps when sending data:

[0034] Step 1: Set a timer when sending data, and store the sent packets in the buffer queue in order;

[0035] Step 2: The timer sets a fixed time. If no confirmation information is received from the receiver within the fixed time period, the sent information is added to the unsent data in the sliding window and waits for the next transmission.

[0036] Furthermore, the workflow of the system includes the following steps:

[0037] Step 1: Use the target marine life video acquisition module to obtain the video stream of the target marine life;

[0038] Step 2: Use the deep learning model training module to update the deep learning model on each drone;

[0039] Step 3: The marine life attribute acquisition module processes the acquired video stream based on the updated deep learning model to obtain the attribute values ​​of the target marine life;

[0040] Step 4: The marine organism attribute encoding module performs feature dynamic differential compression encoding on the attribute change values ​​of the target marine organisms that have changed;

[0041] Step 5: The current drone caches the feature dynamic differential compression code into the queue until the feature dynamic differential compression code in the queue exceeds the preset number of entries, and uses the roaring bitmap compression coding module to perform roaring bitmap compression coding on all the feature dynamic differential compression codes in the queue, and merges them into one message;

[0042] Step 6: Send the message to the buoy to which the current UAV is going to return through the roaring bitmap compression coding sending module;

[0043] Step 7: Determine whether the current drone needs to return to charge. If yes, proceed to step 8; if not, proceed to step 9;

[0044] Step 8: The current drone returns to its buoy to charge, and uses the drone alternation module to forward the received roar bitmap compression encoding information to the new drone that replaces the current drone, and returns to step 4 to continue the monitoring mission of the target marine life;

[0045] Step 9: After receiving the message, the buoy decodes the roar bitmap compression code, converts the code into an underwater acoustic link code through the delay-tolerant relay protocol, and sends it to the monitoring center through the underwater acoustic link. The monitoring center decodes the received underwater acoustic link code.

[0046] Therefore, the present invention adopts the above-mentioned system for marine ecological monitoring information compression and data transmission, which has the following beneficial effects:

[0047] First, the present invention proposes a delay-tolerant relay protocol. In view of the fact that in the entire communication link, the drone and the buoy, and the buoy and the monitoring center cannot obtain each other's status information and cannot perceive the quality of the communication link between the two, in order to improve the success rate of message sending and receiving, a timeout retransmission mechanism is proposed. If the sender does not receive a response from the receiver within a certain period of time, it will assume that the message reception has failed and resend the message; and in order to expand the communication range, the present invention sends the message sent by the drone to the buoy via a wireless link, and the buoy then switches the communication protocol and sends the message to the detection center via the hydroacoustic link, thereby realizing long-distance transmission in the ocean.

[0048] Second, the present invention proposes characteristic dynamic differential compression coding. It is a compression coding designed by utilizing the characteristic that the relevant attributes of the target marine organisms will not change significantly within a period of time. The drone will re-monitor the relevant information of the target marine organisms at fixed time intervals, and compare the monitoring results with the previous monitoring. For the attributes that have not changed much in the results, they will not be sent this time. For the attributes that have changed significantly, the change value of the attribute is encoded, and an index code table is designed to indicate the attribute, so as to reduce the message length and transmission frequency, and reduce bandwidth occupancy.

[0049] Third, the present invention proposes roaring bitmap compression coding. In order to reduce the frequency of sending monitoring messages, the present invention will temporarily store multiple feature dynamic differential compression codes and then send them at one time. In these codes, the change value of each attribute may have the same high bit. For these same high bits, they are only transmitted once, thereby further reducing the length of the message.

[0050] Fourth, the present invention introduces a distributed model update method of federated learning. This method reduces the communication overhead required during training through federated learning, achieves the purpose of model compression through knowledge mutual distillation, better adapts to the characteristics of low computing power of drones, and uses federated asynchronous updates to ensure that drone models can be updated when communication conditions are good, so as to adapt to the unstable communication conditions in the open sea.

[0051] In summary, the present invention is aimed at the marine ecological monitoring business and proposes a system for marine ecological monitoring information compression and data transmission. It uses a federal framework to asynchronously complete the update of each deep learning model in the drone while only transmitting smaller model parameters; and in view of the problems of long ocean communication distance and poor communication conditions, feature differential compression coding and roaring bitmap compression coding are used to reduce the bandwidth overhead in the training and information transmission process, ensuring that the drone can effectively monitor marine life in the open sea environment and successfully send the monitoring results to the monitoring center.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the communication process between drone, buoy and monitoring center;

[0054] Figure 2 Distributed training process for models carried by drones;

[0055] Figure 3 Description of the content sent to the drone;

[0056] Figure 4 It is a description of the feature differential coding frame format;

[0057] Figure 5 Description of the format for the roaring bitmap compression frame;

[0058] Figure 6 This is the flowchart for scenario one;

[0059] Figure 7 This is the signaling diagram for scenario 1;

[0060] Figure 8 This is the flowchart of scenario 2;

[0061] Fig. 9 This is the signaling diagram for scenario 2;

[0062] Fig.10 This is the flowchart for scenario three;

[0063] Fig.11 This is the signaling diagram for scenario three;

[0064] Fig.12 This is a compression ratio comparison chart. DETAILED DESCRIPTION

[0065] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art make various changes or modifications to the present invention, and these equivalent forms also fall within the scope limited by the claims attached to the application.

[0066] The present invention proposes a system for marine ecological monitoring information compression and data transmission, which mainly includes four parts: delay-tolerant relay protocol, feature dynamic differential compression coding, roaring bitmap compression coding, and asynchronous federated model update. The following introduces each part separately:

[0067] 1. Delay-Tolerant Relay Protocol

[0068] First, during the communication process, the sender and the receiver cannot know the specific state of each other, and the channel conditions between the two are also difficult to perceive, so the loss of messages in data transmission is inevitable. In order to ensure the successful reception of messages, the present invention designs a timeout retransmission mechanism, sets a timer when sending data, and stores the sent packets in sequence in a buffer queue. The timer sets a fixed time. If the confirmation information returned by the receiver is not received within this time period, the sent information is added to the unsent data in the sliding window, waiting for the next transmission.

[0069] In addition, in order to expand the transmission range of the message so that the monitoring center can successfully receive the message sent by the drone, the present invention designs a signal relay retuning method. First, the buoy receives the acquisition data packet sent by the drone through a wireless link, and first demodulates the received analog signal through a demodulator; secondly, the demodulated data is transmitted to the buoy deployed at sea for processing, including data decryption processing, data packet merging, decompression processing, verification of data integrity, and error correction processing. Finally, compression and encryption processing are performed, and the processed data is converted into a hydroacoustic signal through a hydroacoustic modulator and sent through a hydroacoustic channel.

[0070] 2. Feature dynamic differential compression coding

[0071] The present invention utilizes the characteristic that the relevant information of the target marine life will not change significantly within a period of time. The drone will re-monitor the target marine life at fixed intervals, use the optical camera carried by the drone itself to capture the relevant video stream, and use multiple deep learning models carried on the drone (for example, target detection, target positioning, speed detection model, etc.) to infer the category and motion information attributes of the target marine life. The inference results of each attribute will be compared with the results of the last message sent, and the change value will be represented by one-hot coding. If the two coding values ​​are different, it is considered that the predetermined threshold is exceeded, indicating that the relevant attribute has undergone a huge change, and the change value of the attribute needs to be encoded into a 0, 1 bit stream for transmission. However, a message may contain the change value encoding of multiple attributes, and the format of the change value after feature dynamic differential compression encoding is as follows: Figure 4 As shown, from Figure 4 It can be seen that the code includes multiple attributes, including drone number, drone position change, marine life number, species category change, current position change, movement speed change, movement direction change, and timestamp field; therefore, an index code table is also designed to identify the names of the attributes represented in the code. This compression coding method can effectively reduce the length and frequency of drone message transmission.

[0072] 3. Roaring bitmap compression encoding

[0073] In order to adapt to the weakly connected communication environment in the open sea, the drone needs to further reduce the frequency of message transmission. Therefore, the drone will first temporarily store multiple feature dynamic differential compression codes, and when the accumulated number reaches 5, it will synthesize them into one message and send them all. In order to further reduce the transmission length of the message, among the multiple feature dynamic differential compression codes, some of the encoded segments for the same attribute have the same high bits. First, the encoded segments with the same high bits in each attribute are extracted. These identical high bits can be extracted and sent only once, and the corresponding low bits are added at the end. In order to facilitate decoding, the present invention also adds the low-bit number and attribute name encoding respectively, such as Figure 5 As shown, for example, the bits corresponding to the two changes of the motion speed attribute are represented as 0111, 0100 respectively. Then the high bit is the first half, that is, the first two bits are 01. The low bit is the second half, which is 11, 00 respectively. In this way, the common high bit and the respective low bits are extracted. The number of low bits is the number of low bits in a common high bit, which is 2 (11, 00) in this example. The attribute name is represented by different 01 bits: for example, 0000 represents speed and 0001 represents x coordinate.

[0074] 4. Asynchronous federated model update

[0075] In the marine environment, the centralized training of the deep learning model carried by the drone in the prior art requires the training data on each drone to be transmitted to the buoy through a wireless link. These data are generally pictures, video streams, etc. The transmission of these data requires a large amount of communication resources, and the training method is difficult to adapt to the offshore environment with poor communication conditions. Therefore, the present invention designs an asynchronous deep learning model update strategy based on a federated architecture. The specific process of using this strategy to update the deep learning model is as follows: first, the personalized model carried on the drone and its corresponding teacher model use the local data collected by the drone for mutual distillation training. Secondly, after the local training is completed, the drone will evaluate the communication link. If it is confirmed that the communication conditions are good, the local model (i.e., the teacher model) data will be transmitted to the buoy through the wireless link; thirdly, the global model generator deployed in the buoy will aggregate the received local model parameters every fixed time period to generate a new global model; finally, when the drone returns to the buoy, the buoy sends the updated global model to the drone, and the drone uses the global model to complete the update of the local teacher model.

[0076] Among them, the personalized models are various deep learning models with the same structure but different model parameters carried by different drones. The type and quantity of deep learning models carried by each drone can be arbitrarily configured according to the task category to be performed by the drone.

[0077] Example

[0078] The present invention is aimed at the following scenarios:

[0079] Scenario 1: The drone needs to monitor the target marine life and send the monitoring data to the monitoring center. In this scenario, the drone needs to use the deep learning model to infer the collected target marine life video stream, obtain relevant results, and then encode the results and send them to the monitoring center through the buoy. The monitoring center decodes the code to restore the data.

[0080] In this scenario, first, the drone captures the video stream of the target marine life through the optical camera. Secondly, the relevant video stream is input into the target recognition model, speed reasoning model, and positioning reasoning model carried by the drone, so as to obtain the species and movement speed attribute results of the target marine life. Thirdly, the drone compares the relevant attribute results with the attribute results recorded when the message was sent last time and generates the corresponding feature dynamic differential compression code. When the number of feature dynamic differential compression codes temporarily stored by the drone reaches the threshold, it is compressed into a roaring bitmap code and sent to the buoy through a wireless link. Then, the buoy decodes the received roaring bitmap code, converts the received 01 bit stream into an integer or floating point number, and converts the code into an underwater acoustic link code through a delay-tolerant relay protocol, and sends it to the monitoring center through the underwater acoustic link. Finally, the monitoring center decodes the received code.

[0081] Scenario 2: When the drone is monitoring and tracking the target marine life, due to energy consumption, the drone needs to return to the buoy for charging. At this time, another drone needs to be activated to continue the monitoring task of the target marine life. In this scenario, the latest monitoring results collected by the returning drone need to be sent to the new drone. First, the returning drone needs to forward the roar bitmap code to the buoy first. Secondly, the buoy forwards the code to the buoy where the drone that is about to take over the task is located through the hydroacoustic link. The buoy then forwards the code to the drone that takes over the task through the wireless link.

[0082] Specifically, in this scenario, first, all the feature dynamic differential compression codes that need to be temporarily stored by the returning UAV are formed into roaring bitmap compression codes through the roaring bitmap compression method, and then the roaring bitmap compression codes are sent to buoy A where the returning UAV is located through a wireless link; secondly, buoy A decodes the received roaring bitmap code, and converts the code into an underwater acoustic link code through a delay-tolerant relay protocol, and then sends it to buoy B where the UAV that is about to take over the task is located through the underwater acoustic link; thirdly, buoy B decodes the received roaring bitmap code, and converts the code into a wireless link code through a delay-tolerant relay protocol and sends it to the UAV that takes over the task; finally, the new UAV decodes and completes the data restoration, thereby replacing the returning UAV to continue to complete the monitoring task of the target marine life.

[0083] Scenario 3: The drone is equipped with different deep learning models to infer different information about the target marine organisms, such as species, movement speed and direction. In this scenario, due to the limitations of offshore communication conditions, the federated learning architecture is difficult to uniformly update the models on each drone at the same time. At this time, the drone needs to asynchronously send the parameters of each model it carries to the buoy based on the communication conditions of the current sea area. Then the buoy completes aggregation and generates a global model based on the received local model parameters at fixed intervals, and then sends the global model to the drone, which then completes the update of the local model based on the global model.

[0084] In this scenario, first, the personalized model carried by the drone and its corresponding local teacher model are trained by mutual distillation; secondly, the drone will choose whether to send the parameters of the local teacher model in real time according to the current communication conditions in the sea area. When the communication conditions are good, the drone will send the model parameters of all local teacher models updated this time to the buoy. The buoy will aggregate the received parameters of each deep learning local teacher model at regular intervals and generate a new global model through the global model generator; finally, when the drone returns to the buoy, the buoy will send the parameters of these global models to the drone, and the corresponding drone will update the corresponding local teacher model according to the received global model parameters.

[0085] In order to verify the feasibility and effectiveness of the method proposed in the present invention, a series of simulation experiments were conducted in this embodiment. In these experiments, 100 relevant data were collected from the simulated unmanned aerial vehicle (UAV), including the target marine life type, target marine life coordinates, target marine life speed, UAV remaining power, engine temperature and UAV coordinates. Table 1 shows the change threshold of each data point compared with the previous transmission. In the experiment, the data to be transmitted was randomly generated.

[0086] Table 1 Data change threshold

[0087] property Threshold Target marine life types Type changes Target marine life coordinates More than 2km away from the last coordinate location Target marine life speed 4km / h The remaining battery power of the UAV 2% UAV engine temperature 30℃ Coordinates of the UAV More than 1km away from the last coordinate location

[0088] In this example, the following four different propagation modes were evaluated:

[0089] The first method is to transmit raw data without any compression processing. The drone transmits every data point it collects; the second method is dynamic index table encoding. Only data with changes exceeding a threshold is transmitted, and there is an index before the data string to identify the relevant content; the third method uses bitmap compression encoding. Once the number of changes reaches a threshold, a batch of changes will be temporarily stored and transmitted; the fourth method uses the encoding method proposed by the present invention.

[0090] In terms of compression efficiency, this embodiment mainly evaluates the compression performance of the message length and message transmission frequency of the four methods. The message length transmitted after each data aggregation and the amount of messages transmitted by each method are compared to measure their compression efficiency. In the experiment, the threshold of the amount of messages stored instantly by the drone is set to 5. This means that when the number of temporarily spare messages reaches 5, the drone just transmits information.

[0091] according to Fig.12 By calculating the number of rounds and data length, it can be obtained that dynamic index table encoding, bitmap compression encoding, Huffman encoding and this application achieved message length compression rates of 14.5%, 61.3%, 85.9% and 9.3% respectively. The message sending frequency compression rates are 61%, 21%, 100% and 12% respectively relative to the original data transmission method. These results show that the compression scheme proposed in the present invention can fully reduce the message length and repetition frequency of message transmission of drones.

[0092] Fig.12 It is shown that bitmap compression coding has the most astounding message length compression ratio, because it stores various changes and conveys the stored information together, radically shortening the message length. Dynamic index table coding has the most noteworthy recursive pressure ratio, because it only conveys the changed information on the side, greatly reducing the amount of transmitted messages. Compared with these baseline strategies, the strategy proposed by the present invention achieves a stricter message length and repetition pressure ratio.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A system for marine ecological monitoring information compression and data transmission, comprising multiple drones and multiple buoys deployed in different sea areas, characterized in that: Also includes: A target marine life video acquisition module is used to acquire a video stream of the target marine life through an optical device on a group of drones; A deep learning model training module, which is used to update the deep learning models on different drones using a federated asynchronous training method; The marine life attribute acquisition module is used to obtain the attribute values ​​of the target marine life through different deep learning models on the drone; The marine life attribute coding module is used to compare the current value of each marine life attribute with the attribute value obtained last time. If there is a change, the change value of the attribute is subjected to characteristic dynamic differential compression coding; The roaring bitmap compression coding module is used to perform roaring bitmap compression on multiple feature dynamic differential compression codes stored in the same queue to form a message and send it to the roaring bitmap compression coding sending module; The roaring bitmap compression encoding sending module is used to send the received message to the buoy to which the UAV is to return; The drone alternation module is used to replace the current drone with a new one when the drone currently performing the mission needs to return for charging, and continue to perform the monitoring mission of the target marine life.

2. A system for marine ecological monitoring information compression and data transmission as claimed in claim 1, characterized in that: The federated asynchronous training method in the deep learning model training module comprises the following steps: Step 1: Each drone uses the local dataset to perform mutual distillation training on the deep learning model it carries and the local teacher model corresponding to the deep learning model; Step 2: Each UAV obtains the communication status of the current sea area and asynchronously sends the parameters of each model carried by the UAV to the buoy to which it is returning; Step 3: The global model generator deployed in the buoy aggregates the received model parameters at fixed time intervals to generate a new global model; Step 4: When the UAV returns to the corresponding buoy, the buoy sends the updated global model to the UAV, and the UAV uses the global model to update the local teacher model.

3. A system for marine ecological monitoring information compression and data transmission as claimed in claim 1, characterized in that: The attributes of the target marine organism include the category of the target marine organism, the moving speed of the target marine organism, the current position of the target marine organism, and the moving direction of the target marine organism.

4. A system for marine ecological monitoring information compression and data transmission as claimed in claim 3, characterized in that: The marine life attribute coding module includes an attribute change comparison unit and a feature dynamic difference compression coding unit; The attribute change comparison unit is used to compare the attribute values ​​of the current target marine organisms with the corresponding attribute values ​​last sent to the buoy. If the attribute change value exceeds the preset threshold value of the attribute, the attribute change value is sent to the characteristic dynamic differential compression coding unit. The characteristic dynamic differential compression coding unit is used to perform binary coding on the change value of the attribute and add index coding to obtain the characteristic dynamic differential compression coding.

5. A system for marine ecological monitoring information compression and data transmission as claimed in claim 4, characterized in that: The frame format of the characteristic dynamic differential compression encoding includes a drone number field, a drone position change field, a marine life number field, a species category change field, a current position change field, a movement speed change field, a movement direction change field and a timestamp field in sequence, and the name of each attribute field is identified by an index code table.

6. A system for marine ecological monitoring information compression and data transmission as claimed in claim 5, characterized in that: The steps of roaring bitmap compression coding module for roaring bitmap compression coding are: first extract the coding segment with the same high bit in the characteristic dynamic differential compression coding of each attribute, and add the corresponding low bit and the number of low bits and attribute name coding after the coding segment to form the final roaring bitmap compression coding.

7. A system for marine ecological monitoring information compression and data transmission as claimed in claim 6, characterized in that: The specific steps of the drone alternation module using a new drone to replace the current drone to continue the mission are: Step 1: Perform roaring bitmap compression on all feature dynamic differential compression codes cached by the returning drone to obtain roaring bitmap compression code; Step 2: The returning drone sends the roar bitmap compression code to buoy A where the returning drone is located through a wireless link; Step 3: Buoy A decodes the received roar bitmap code, converts the code into an underwater acoustic link code through a delay-tolerant relay protocol, and then sends it to buoy B where the new UAV that is about to take over the mission is located through the underwater acoustic link; Step 4: Buoy B decodes the received roar bitmap code, converts the code into a wireless link code through a delay-tolerant relay protocol, and then sends it to the new UAV; Step 5: The new drone decodes and restores the data, thereby replacing the returning drone to continue the monitoring mission of the target marine life.

8. A system for marine ecological monitoring information compression and data transmission as claimed in claim 7, characterized in that: The delay-tolerant relay protocol includes the following steps when sending data: Step 1: Set a timer when sending data, and store the sent packets in the buffer queue in order; Step 2: The timer sets a fixed time. If no confirmation information is received from the receiver within the fixed time period, the sent information is added to the unsent data in the sliding window and waits for the next transmission.

9. A system for marine ecological monitoring information compression and data transmission as claimed in claim 7, characterized in that: The workflow of the system includes the following steps: Step 1: Use the target marine life video acquisition module to obtain the video stream of the target marine life; Step 2: Use the deep learning model training module to update the deep learning model on each drone; Step 3: The marine life attribute acquisition module processes the acquired video stream based on the updated deep learning model to obtain the attribute values ​​of the target marine life; Step 4: The marine organism attribute encoding module performs feature dynamic differential compression encoding on the attribute change values ​​of the target marine organisms that have changed; Step 5: The current drone caches the feature dynamic differential compression code into the queue until the feature dynamic differential compression code in the queue exceeds the preset number of entries, and uses the roaring bitmap compression coding module to perform roaring bitmap compression coding on all the feature dynamic differential compression codes in the queue and merge them into one message; Step 6: Send the message to the buoy to which the current UAV is going to return through the roaring bitmap compression coding sending module; Step 7: Determine whether the current drone needs to return to charge. If yes, proceed to step 8; if not, proceed to step 9; Step 8: The current drone returns to the buoy to charge, and uses the drone alternation module to forward the received roar bitmap compression encoding information to the new drone that replaces the current drone, and returns to step 4 to continue the monitoring mission of the target marine life; Step 9: After receiving the message, the buoy decodes the roar bitmap compression code, converts the code into an underwater acoustic link code through the delay-tolerant relay protocol, and sends it to the monitoring center through the underwater acoustic link. The monitoring center decodes the received underwater acoustic link code.

Citation Information

Cited By

  • Lightweight ocean data acquisition and processing system and method based on edge calculation

    CN120238585A